In a joint technical study published on arXiv and an official release by Anthropic, scientists from Harvard University, Anthropic, and QuEra Computing have introduced an evolutionary AI framework where Large Language Models (LLMs) automatically discover, write, and optimize fault-tolerant quantum circuits.
Detailed in the paper (“Discovering state-of-the-art quantum circuits with large language models“) and Anthropic’s technical release (“Designing more efficient quantum circuits with Claude“), the research demonstrates that pairing LLM code proposals with an independent classical evaluator reduces the space-time resource costs (T × Q-cost) of core computational primitives by 4.1x to over 9,100x relative to hand-crafted baselines.
| [ Resource Improvements Across Eight Compilation Subroutines ] | |||
|---|---|---|---|
| Compilation Subroutine / Task | Baseline (T, Q) | Discovered (T, Q) | TQ-Cost Gain (G_TQ) |
| • Reed–Solomon DQI Decoder | • (5.1 × 106, 1,938) | • (4.4 × 105, 1,837) | • 12x Reduction |
| • Hermitian DQI Decoder | • (1.0 × 108, 1.9 × 105) | • (9.3 × 105, 2,395) | • > 9,100x Reduction |
| • HHL Rotation Angle (arcsin) | • (1.6 × 104, 300) | • (1,348, 115) | • 32x Reduction |
| • Coulomb Potential (1/√||x||²) | • (6,355, 252) | • (1,140, 142) | • 9.9x Reduction |
| • Fixed-Point Gaussian (e-x²) | • (2.4 × 104, 362) | • (1,583, 139) | • 39x Reduction |
| • Fe₂S₂; Alias–Select (Chemistry) | • (5.5 × 104, 139) | • (1.5 × 104, 126) | • 4.1x Reduction |
| • Non-Abelian Lattice-Gauge Sim. | • (1.3 × 106, 1,927) | • (1.7 × 105, 1,859) | • 7.7x Reduction |
| • Prime-Field Modular Inversion | • (1.9 × 106, 1,446) | • (4.0 × 106, 935) | • New Divsteps Architecture |
Evolutionary AI Framework and Closed-Loop Co-Design
Rather than relying on local gate-level rewrites, the system deployed frontier LLM models (Claude Mythos 5) in an automated closed loop: proposing complete Python programs across distinct algorithmic families, evaluating basis trajectories in a classical simulator (kickmix model), receiving feedback on Toffoli gate counts and peak qubit footprints, and autonomously refining circuit architectures.
Highlighting the impact of algorithmic optimization over pure hardware scaling, the research notes that classical estimates for simulating the FeMoco iron–sulfur cluster dropped from 50 trillion Toffoli gates in 2017 to ~340 million through circuit redesigns—a ~150,000-fold reduction achieved over eight years of human research. In contrast, the AI framework arrived at a new reversible modular inversion architecture in a single 53-hour autonomous run with minimal human intervention.
The evolutionary search repeatedly uncovered three fundamental co-design principles:
- Conditional Routing over Conditional Arithmetic: Replacing separate, expensive conditional arithmetic blocks for every execution branch with inexpensive controlled SWAP gates that route selected data through a single shared, uncontrolled arithmetic block.
- Branch Decoding from Evolving Arithmetic State: Eliminating massive qubit history stacks by decoding intermediate branch choices directly from mathematical invariants present in the updated arithmetic registers during the uncomputation pass.
- Bounded Resource Allocation via Failure Budgets: Allocating history stack space and loop bounds to match typical inputs rather than worst-case limits, permitting a controlled failure rate (ε ~ 10-4) to save hundreds of ancilla qubits.
Modular Inversion, Formal Proofs, and DQI Decoding
In the eighth task—prime-field modular inversion for elliptic-curve cryptography—the framework discovered a new reversible architecture for divsteps, a classical constant-time modular inversion algorithm previously considered unusable in quantum circuits due to history storage overhead. Rather than storing full branch histories, the AI derived a 2-adic mathematical formula that extracts intermediate branch choices directly from the updated registers, verifying the proof in the Lean theorem prover. By decoding 87.6% of ambiguous branch decisions and storing only 12.4% on a 60-qubit stack, the circuit established a new Pareto frontier for fault-tolerant point addition.
For the Hermitian code Decoded quantum interferometry (DQI) decoder (>9,100x savings), the AI replaced raw null-space matrix storage with an in-place Lanczos recurrence and a striped phase scan, dropping the peak qubit count from 190,000 to 2,395 and Toffoli gates from 100 million to 930,000. For Reed–Solomon decoders, switching from the extended Euclidean algorithm to Berlekamp–Massey decoding delivered a 12x reduction in Toffoli gates.
Review Yuval Boger’s post on LinkedIn here, examine the research feature on Anthropic Research here,and inspect the technical manuscript on arXiv here.
October 17, 2026